wi m ab gr DL Mh hesd ut SR “AH 
“pe A ih “HAF TL Re = EE 
GDS Sle ae ARTIC eR Bad. 
CE gy RG 2 Yo hel ed Dodge $ 
gd ES amg CL eh kT Boer 
NL] a CH 
ce EAT 4 A $i we Sede TY 
YEN, ep BIEN Ds E EN BUTE ee 
Cia vl Bd, dy dh vB eT 
Cad eR RT NE Tet 
(a) Brightness level 6 — 684 objects (b) Brightness level 7.7 — 3516 objects 
Figure 2: Effect of chosen brightness of the microscope on object number. Both images were 
binarised with threshold 160. 
- an iterative algorithm originally for maximum likelihood estimation of the parameters of a 
multivariate (Poisson) distribution, see e.g. [LR87] for details. The estimated parameters are 
interpreted as weights deciding how strongly the pixels in each bin contribute to the content 
of A-, D-, or E-graphite, respectively. 
2.3 Automatic classification. After the learning phase the algorithm is able to classify 
automatically but still reflecting the typical style of classification of the teacher. Figure 3 shows 
ral tothe two classified images. 
do 3 The image sample 
ırf. and [TWM Quality and range of the image sample used for teaching are crucial for the success of the 
algorithm. Shading, over-polishing, and impurities (nonmetallic inclusions or small dust par- 
ticles) should be avoided. While shading correction could be part of the algorithm it is nearly 
impossible to correct for the over-polishing and impurities in a robust way. 
Another source of instability is the binarisation. The finer the structures are the more the 
far grey cast Iron result is affected by the binarisation. This could be avoided by using methods based on the 
je interDretation distance transformation. 
able See Figure An important requirement is that the teaching sample should span the whole range of 
je Dicture — the images to be classified by the algorithm. That is, the sample should represent the production 
range of the foundry. If, for example, the teaching sample does not include any A-graphite 
operations then the algorithm will not be able to detect A-graphite. Pure samples, that is samples with 
ınd or {he union 100% A-, D-, or E-graphite, seem to be good teaching examples because there is high certainty 
+ gz directions in the manual classification. However, experiments show that they are too “ideal” to teach the 
N x geometric classification of composite structures. Further experiments would be needed to determine the 
„ hins according optimal mixture of pure and composite structures in the teaching sample. 
4 Evaluation and comparison 
wo wr Y It is difficult to evaluate the automatic classification as it is difficult to compare different 
. Cnr classifications. Here the mean distance, normalized such that its range is [0, 1], for a validation 
185